You are on page 1of 14

ESSAY

www.advintellsyst.com

From Bit to Bedside: A Practical Framework for Artificial


Intelligence Product Development in Healthcare
David Higgins* and Vince I. Madai

treatment, costs are increasing enor-


Artificial intelligence (AI) in healthcare holds great potential to expand access to mously. Thus, there is an urgent need to
high-quality medical care, while reducing systemic costs. Despite hitting head- deliver better quality healthcare while
lines regularly and many publications of proofs-of-concept, certified products are simultaneously lowering costs to keep our
healthcare systems sustainable.
failing to break through to the clinic. AI in healthcare is a multiparty process with
Technological innovation appears to be
deep knowledge required in multiple individual domains. A lack of understanding the only mechanism which has the poten-
of the specific challenges in the domain is the major contributor to the failure to tial to fulfill both of these, seemingly
deliver on the big promises. Herein, a “decision perspective” framework for the contradictory, requirements. Artificial intel-
development of AI-driven biomedical products from conception to market launch ligence (AI), more specifically machine
learning (ML)-based AI, has become one of
is presented. The framework highlights the risks, objectives, and key results
the main technologies driving the so-called
which are typically required to navigate a three-phase process to market-launch fourth industrial revolution.[3] The main ML
of a validated medical AI product. Clinical validation, regulatory affairs, data method at the forefront are so-called deep
strategy, and algorithmic development are addressed. The development process artificial neural networks (ANNs). ANNs
proposed for AI in healthcare software strongly diverges from modern consumer are inspired by simplified neuronal struc-
software development processes. Key time points to guide founders, investors, tures and consist of layers of artificial
neurons. The strengths of the connection
and key stakeholders throughout the process are highlighted. This framework
between neurons of different layers are
should be seen as a template for innovation frameworks, which can be used to determined in the training process and
coordinate team communications and responsibilities toward a viable product determine the predictive capability of the
development roadmap, thus unlocking the potential of AI in medicine. model.[4] ANNs are general function approx-
imators, in theory, capable of approximating
any mathematical function of data.[5] Other
non-ANN ML methods successfully applied
1. Introduction in healthcare include decision tree algorithms,[6–8] generalized
linear models,[9] support vector machines,[10,11] and Gaussian pro-
Healthcare systems all over the world face tremendous chal- cesses (GPs).[12,13] Modern ML methods constitute the new state
lenges. The age-related illness burden is increasing, particularly of the art in computer vision,[14] natural language processing,[15]
in wealthy countries, due to ageing populations. Lifetime risk for and recommender systems[16–18] facilitating technologies from
cancer has reached up to 50%[1] and lifetime risk for stroke is smart assistants to self-driving cars. Increasingly, they are applied
25%.[2] As a consequence of the increasing incidence of these for healthcare use cases. They have the potential to deliver person-
costly diseases, along with broadening access to healthcare alized treatments[19–21] and monitoring,[22,23] with lower error
and major advances in pharmaceutical and technological disease rates,[23,24] at greatly reduced costs.[20,23]
Many publications share the vision that AI in healthcare will
achieve the aforementioned goals to keep our healthcare systems
Dr. D. Higgins sustainable.[19,20,25–27] However, the overwhelming majority of
Berlin Institute of Health
Anna-Louisa-Karchstr. 2, 10178 Berlin, Germany reported positive studies to date are retrospective in nature and
E-mail: dave@uiginn.com the presented solutions, more often than not, provide mere proofs
Dr. V. I. Madai of concept (PoC) as evidenced by systematic reviews.[28,29] AI/ML-
Charité Lab for Artificial Intelligence in Medicine (CLAIM) based tools which are applicable in the clinical setting—with medi-
Charité – Universitätsmedizin Berlin cal device certification, clinical validation, and routine clinical use—
Augustenburger Platz 1, 13353 Berlin, Germany are still scarce. Despite a rapidly increasing number of scientific
The ORCID identification number(s) for the author(s) of this article publications describing potential applications, products benefiting
can be found under https://doi.org/10.1002/aisy.202000052. patients and our healthcare systems are not deployed at the rates
© 2020 The Authors. Published by WILEY-VCH Verlag GmbH & Co. KGaA, necessary. This translational gap constitutes a major public health
Weinheim. This is an open access article under the terms of the Creative challenge. One of the most promising technologies of recent years
Commons Attribution License, which permits use, distribution and is not reaching the patients and healthcare systems in need.
reproduction in any medium, provided the original work is properly cited. As the translational gap refers to a block on transforming PoC
DOI: 10.1002/aisy.202000052 into actual products, we are facing a crisis of AI in healthcare

Adv. Intell. Syst. 2020, 2, 2000052 2000052 (1 of 14) © 2020 The Authors. Published by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
www.advancedsciencenews.com www.advintellsyst.com

product development. Thus, in this work, we outline a novel David Higgins is a mathematician and
framework of AI in healthcare product development. Product computer scientist with a Ph.D. in
development is defined as “the transformation of a market computational neuroscience. For the
opportunity and a set of assumptions about product technology past 2 years, he has led industrial
into a product available for sale.”[30] It is possible to focus on research projects in pharma, and built
product development from different angles, i.e., marketing, orga- behavioral models for patient
nizational structure, engineering, and operations. However, interventions via consumer
these views are highly specialized and have a tendency to over- technology. Currently, he is a visiting
or underfocus on certain challenges. We base our framework on researcher at the Berlin Institute of
the model of the “decision perspective” where a string of deci- Health. He is mentor to a number of
sions transforms an idea into a product.[30] This model is based AI-driven health-tech spin-outs.
on the observation that “what is being decided” is fairly consis-
tent over the course of all types of product development, whereas Vince I. Madai is a medical AI
how decisions are implemented can vary tremendously.[30] Thus, researcher at Charité Berlin and the
importantly, our framework does not give many recommenda- scientific lead of the “Charité Lab for
tions on how to implement the decisions that are suggested. Artificial Intelligence in Medicine -
We focus largely on what needs to be decided for a certain stage CLAIM.” He is an M.D. with a Ph.D.
in development to proceed successfully to the next stage and what in medical neuroscience and M.A. in
needs to be decided to keep the risks of failure for further stages medical ethics. As a strong supporter
as low as possible. of interdisciplinary research, his main
The main finding that underlies the presented product interests are the translation of
development framework is that, in contrast to other software machine learning applications into the
products, healthcare-specific differences exist in the develop- clinical setting and the challenge of biased AI tools.
ment path of AI healthcare products. Product development
for a typical technology start-up is long and can be broken down
into multiple stages.[31] Typically there is an initial “formation
stage,” in which a team comes together around a potential On the other hand, the framework allows scientists, regulators,
product idea. This is followed by a “validation stage,” in which investors, and political actors to better understand the specific
prototypes are turned into products and an eventual business needs of AI in healthcare projects and enterprises.
model is settled-on. Also, finally, there is a “scaling phase,”
where the addition of considerable capital allows the small
company, with a now validated product–market fit, to quickly 2. Results
expand to capture considerable market share of an existing mar-
We have developed a framework following three consecutive
ket or to expand into a newly discovered market niche. In this
phases and covering four domains for AI/ML-driven clinical
approach, current methodologies drive entrepreneurs toward
product development. The three phases follow an entrepreneur-
early contact with potential customers and early trials of incom-
ial timeline. We break the process down into three distinct prod-
plete versions of the final product.[32] For AI products aimed at
uct development phases before a product exists in the market.
healthcare, the main hidden difficulty is that to capture the
Each of these phases is self-contained and typically can be recog-
enhanced value available for medical treatments evidence must
nized by both investors and regulators as a distinct stage on the
first be presented that the product is safe. In addition, to be
road to market. For each of the phases, our framework spans
incorporated into standards of care, the product must also be
across four major domains which are relevant for the develop-
shown to either enhance treatment outcomes or decrease costs.
ment of an AI/ML product in healthcare: clinical, regulatory,
This requires a validation pipeline tailored to regulatory and
data, and model development. Finally, we present an overview
clinical requirements[33] which make data acquisition and anal-
of the most important postmarket risks, delivering software
ysis time-consuming and difficult. Current software product
updates, changes in medical practice, and surveillance, each of
development practices are rarely a fit for both the medical
which must be mitigated against once the product enters the
certification as well as the clinical validation process. This—
market.
in turn—leads to 1) much longer development cycles and
In the following, we will give definitions of the three phases
2) quite different proof points inherent in developing a product
and four domains before outlining the framework in detail. The
which must be medically certified.
phases and domains are shown in Figure 1, 2, 3, and 4 following
Thus, we argue that to develop an AI in healthcare product,
a domain perspective.
the entire development cycle must be changed. We present a
rethought product development cycle, primarily aimed at prod-
ucts in the clinical setting, bridging the cognitive gap between 2.1. Definition of the Three Phases
modern digital development practices and current biotech prac-
tices. The purpose of this map is, on one hand, to enable poten- Phase 1, called “Form”, involves bringing together a small
tial digital health founders in the area of AI in healthcare to have team/group around a potential clinical need, with the purpose
a more accurate perspective of the road ahead, and to allow exist- of providing a preliminary technical solution. This phase may
ing digital health entrepreneurs to better allocate their resources. take place within the support structures of a hospital spin-out

Adv. Intell. Syst. 2020, 2, 2000052 2000052 (2 of 14) © 2020 The Authors. Published by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
www.advancedsciencenews.com www.advintellsyst.com

Figure 1. The clinical domain refers to identifying real-world clinical needs and validating these needs throughout the life cycle of the project. Herein, the
major risks, objectives and key results, and practical advice, across the three time-phases of development, are presented.

unit or company incubator. In this phase, the main goals are to help steer the development process. The goal of this phase is to
to form, and maintain, the group membership, check the tech- build a basic implementation appropriate from a clinical and
nical feasibility of a solution, and to understand the pathway to regulatory point of view. The outcomes of this phase are multi-
market by validating the clinical need. This latter point, we will modal, but the intrinsic value at its end-point is a solid codebase
explain, actually already requires a basic understanding of regu- which demonstrably solves a clinical need.
latory and clinical validation paths. At the end of the Form phase, Phase 3, called “Launch,” is the phase which least matches
the intrinsic value will typically amount to a PoC solution and current (consumer-oriented) software development practices.
little more. A medical product cannot directly enter the market without first
Phase 2 is called the “Build” phase. At this point, a more passing certification and clinical validation steps. Therefore, pro-
defined team decides to work together for a period ranging from cesses such as medical device certification or a clinical study
18 months to 5 years. Clinical members may initially maintain demonstrating efficacy must be conducted. To conduct certifica-
other commitments, but they need to provide enough availability tion, or a study, typically the software version must be frozen.

Adv. Intell. Syst. 2020, 2, 2000052 2000052 (3 of 14) © 2020 The Authors. Published by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
www.advancedsciencenews.com www.advintellsyst.com

Figure 2. Medical device regulation encompasses all laws and rules with regards to the development of a healthcare product falling under the definition of
a medical device. Herein, the major risks, objectives and key results, and practical advice, across the three time-phases of development, are presented.

Adv. Intell. Syst. 2020, 2, 2000052 2000052 (4 of 14) © 2020 The Authors. Published by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
www.advancedsciencenews.com www.advintellsyst.com

Figure 3. For each of the three phases, it is highlighted which data sources founders and developers should focus on. Again, the major risks, objectives
and key results, and practical advice, across the three time-phases of development, are given.

This does not mean, however, that the team stops working. the certification and validation steps. The end result of this phase
In our framework, there is still considerable technical work in is a product which can be deployed, in the majority of appropriate
this period. That said, the main goal of the phase is to complete clinical settings, in the market.

Adv. Intell. Syst. 2020, 2, 2000052 2000052 (5 of 14) © 2020 The Authors. Published by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
www.advancedsciencenews.com www.advintellsyst.com

Figure 4. At every stage of algorithm development, we try to simplify this process outlining the long- and short-term priorities and how the two
should align. This is presented under the framework of major risks, objectives and key results, and practical advice, across the three time-phases
of development.

2.2. Definition of the Four Domains clinical partners, including patients and payers. We visualize,
including supplementary advice, this domain in Figure 1.
2.2.1. Clinical Domain

The clinical domain refers to the process of identifying real-world 2.2.2. Regulatory Domain
clinical needs and validating these needs throughout the life cycle
of the project. Clinical needs are not just medical needs but Medical device regulation encompasses all laws and rules with
extend to identifying different interaction requirements of differ- regards to the development of a healthcare product falling under
ent types of users (UI/UX research), and process needs of all the definition of a medical device. AI in healthcare products are

Adv. Intell. Syst. 2020, 2, 2000052 2000052 (6 of 14) © 2020 The Authors. Published by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
www.advancedsciencenews.com www.advintellsyst.com

highly likely to fall under such regulations. We will explore for all benefit—thus, doctors will be led to use it despite potentially
three phases of a healthcare AI project, how to always be pre- higher treatment costs—or it must reduce costs. If the solution
pared for the regulatory requirements of the next phase and neither reduces costs nor shows a proven clinical benefit, then
to deleverage critical risks early in the project. As the particulars the potential product is not financially viable. The team which
of regulatory processes differ enormously throughout the world, comes together in the Form phase will thus need to be able
we will give only general advice which should apply across to identify the clinical need, to validate it and build a first tech-
regions. A visualization for the regulatory domain appears in nical solution to address this need.
Figure 2.

2.3.2. Regulatory Domain


2.2.3. Data Domain
Regulatory requirements are a significant hidden cost of devel-
Because AI methods are inherently data-driven, the right kind of
opment of medical AI solutions. Medical device regulators, such
data is crucial to develop an AI-based healthcare product. In the
as the Federal Drug Administration (FDA) in the USA and
data domain sections, we will explore which kind of data needs to
European Medicines Agency (EMA) in the EU, have a mandate
be leveraged in each phase. The data necessary for an early PoC
to protect the public from potentially dangerous medical inter-
to get traction can differ tremendously from the data which are
ventions. This means that the choice to be regulated is rarely,
necessary in later stages for certification and for clinical valida-
in fact, a choice. Given this situation, entrepreneurs need to
tion. We will highlight for each of the three phases, what data
sources founders and developers should focus on. Figure 3 understand very early how their future product will be regulated.
shows an overview. In the Form phase, it is unlikely that software developers will
need to follow best practices for the development of a medical
device. Most projects fall apart long before even a pilot study,
2.2.4. Algorithm Domain
so spending time and effort on regulatory compliant software-
development processes at this early stage is inefficient. This is
Another crucial domain is the choice of the AI methods applied
the time, however, for the founding team to familiarize them-
for product development. With regards to the algorithms, atten-
selves with the regulatory road ahead and to be aware of future
tion must be paid to the number of addressable patients in both
requirements. Here, it is prudent to seek professional help as
the early training phases, also in subsequent deployment; to
regulatory requirements are legally complex. This knowledge will
questions of engineering around energy use, timeliness, and
build confidence with potential funders and will deleverage a
location of the data; to neglected populations in biased training
large portion of Build (phase 2) risk ahead of time.
samples; and to the use of algorithms which will support any
planned subsequent integration of other data streams. At every
stage, we try to simplify this process outlining the long- and 2.3.3. Data Domain
short-term priorities and how the two should align. Figure 4
shows the algorithm domain results. New projects frequently lack access to fitting data or realize (too)
late that access to promised data cannot be granted. Many clini-
2.3. Form Phase cians are not aware of legal impediments to share data, for com-
mercial use, so the decision to form a team should be based on
The Form phase of a medical AI project can last 6–12 months clear evidence that data will be available and usable.
and involves the earliest stages of proofing a concept around a When such initial data are successfully secured, it is para-
digital solution to a perceived clinical need. mount to evaluate whether the data are supportive toward solving
the identified clinical need. Here, clinical founders often overes-
2.3.1. Clinical Domain timate the value of their initial data set. Although it is certainly
admirable, and indeed necessary, to bring data to the table at the
The most important goals of the Form phase are to define a clin- outset of a medical AI project, the iterative procedure of product
ical need and to show that it is solvable in the form of a PoC. development will often make the utility of the initial data set
Clinicians are expert practitioners, and as such rely on their redundant.
expertise to identify those clinical needs. Decades of experience Any initial data set should be seen as an early research tool,
in product design and development have shown, however, that from which to draw insights, which will contribute to the product
while expert vision is useful for the identification of product– development process. In the Form phase, it is enough to build
market fit, experts might fall too quickly in love with their own models which can only predict outcomes based on this, typically,
solutions and their own pet pain points.[34,35] This means, that single-sourced—with multiple hidden biases—data. Practitioners
the correct—objective—identification and validation of a clinical should, however, begin to test the limits of their assumptions
need is a high priority for the Form phase of development. The about these data. The urge to cherry-pick and to oversell the data
biggest risk for the Form phase is to focus on work that has no utility, at this point, is only storing up larger problems for later.
validated clinical need. Be aware that a thorough validation of data hypotheses will be
A valid clinical need is not solely defined as a clinical problem unavoidable in later stages. It is thus crucial to come up with a
which needs solving. As development is the goal, the solution for data plan outlining the usefulness of the currently used data
the clinical need at hand must either provide a proven clinical and defining the need for future data sets.

Adv. Intell. Syst. 2020, 2, 2000052 2000052 (7 of 14) © 2020 The Authors. Published by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
www.advancedsciencenews.com www.advintellsyst.com

2.3.4. Algorithm Domain Build phase, it is crucial to follow a predefined and continuously
updated clinical validation plan.
Similar to the data domain, models developed in the Form phase
will often be superseeded in later stages. What is important in 2.4.2. Regulatory Domain
this early phase is to choose algorithms which are fit for purpose,
not only with this smaller data set but also with data on the The Build phase involves an earnest attempt to develop software
intended deployment scale. Testing of multiple algorithms which will form the bedrock of a future product. As such, this is
and the development of performance benchmarks are also very the kick-off point for the development of a solution which will
important. Choosing methods which overperform on the initial withstand regulatory approval. There is the evident risk that
data set, but which store up problems for subsequent expansion, incorrect assumptions are made about the required certification
along with performance metrics which are inappropriate for clin- levels and the associated costs and timelines. In the worst case,
ical applications are common errors at this point. this discovery is made so late in the Build phase that mistakes
A tempting self-deception is to fall into the habit of adopting cannot be undone and the project or startup fails. It is also impor-
bad scientific practices to produce initial results which in turn tant to make sure that every single piece of technology is used in
lead to very high chances of failure in later stages. We understand the correct manner for the certification process.
the need to show quickly initial results which will be the basis to There is a strong trade-off between instituting correct medical
attract further funding—be it public or private. However, sloppy device (MD) certifiable software development processes early ver-
development in this stage will lead to high failure rates in the sus late in the product development cycle. The process overhead
Build phase, even if a clinical need exists and the team has for developing an MD-certified product is considerable, especially
the right data to solve it. in a product which is not yet tightly designed to fit a product–
market niche. However, introducing retrospective audits and
2.3.5. Outlook new development processes late in the development cycle can
be extremely financially costly and time-consuming. Potentially
Form phase is a crucial phase where a few vital key objectives tipping the balance in favor of introducing MD certifiable guide-
need to be kept in mind. Foremost, the focus must lie on the lines early in the process is the fact that this will also bed-down
validation of a clinical need and the development of a PoC to appropriate practices in the development team. Because these
solve it. The clinical need must lead to reimbursement—either practices will need to be continued for the entire lifecycle of
through proven clinical benefit or proven cost reduction. The the product, it makes sense to begin as early as feasible. It is clear,
formalization of the data and algorithmic hypotheses, and asso- however, that for some products, a later start might be the better
ciated benchmarks, is a demanding process for smaller teams. option. The key is that decision-makers actively make an
Thus, it is essential to focus on the key objectives. informed decision as to how to proceed and when development
will switch to the scrutiny required for the certification process.
2.4. Build Phase
2.4.3. Data Domain
Following the early incubation of the Form phase, comes Build
phase, which can take as long as 2–5 years depending on the type This is the stage at which external data sources must be aggre-
of product being developed. In this phase, a concerted effort is gated with internal ones, allowing hypotheses about homogene-
made to develop an AI-driven solution, while also demonstrating ity and data quality to be assessed systematically for the first time.
a clear clinical need. This, however, requires a strong focus on data processing and
harmonization. It is important to make sure that the team has
2.4.1. Clinical Domain not only ML expertise but is also experienced in correct data proc-
essing. Because any work conducted in the Build phase is typi-
During the Build phase, it is highly important to continuously cally in very close partnership with clinics, or partners, it allows a
validate and fine-tune the clinical need, to make sure that the relatively tight feedback loop on data which will not exist in later
development process follows the right path. Due to the multi- stages of development. Here, new data sets allow continuous
modal nature of product development, a typical product will assessment of the model performance. It is crucial to emulate
morph multiple times during this phase. A strong risk, in this in the Build phase the heterogeneity of the real world, e.g., in
phase, is that one ends up with a technical solution which works terms of different hardware or different clinical practices.
well but no longer addresses the actual clinical need. Another Depending on the clinical need, it might be necessary to run a
potential risk is that the solution works on the available data data collection study in the Build phase. For example, data from a
but is actually nonperformant on data from other sources new study might be necessary to reflect a change in current clin-
(i.e., a highly biased model is developed). Thus, it is important ical practice, where retrospective data might not exactly reflect
to work with partner clinics, or practitioners, which are not the use case anymore. It is also possible that in the Form phase
directly under the control of the project initiator. This is partly the PoC was purely technical and data for the specific use case
to avoid complications of medical hierarchy, but also to expose was not, at that point, available. It is crucial to understand that
the product development team to differences in clinical practices, the data set, on which the medical AI product is built, can rarely
hardware, and opinions. Only this exposure will allow the team be used for the subsequent validation of the product itself
to fine-tune the solution to real-world needs. Especially in the (see Section 2.5 and 3). At the end of the Build phase, the data

Adv. Intell. Syst. 2020, 2, 2000052 2000052 (8 of 14) © 2020 The Authors. Published by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
www.advancedsciencenews.com www.advintellsyst.com

set must one-to-one reflect the real-world conditions to allow the early results suggested. The largest effort, in this phase, must be
development of models which will perform as well during medi- applied to bringing all the components together. This applies on
cal certification and clinical validation as in-house. the software side, requiring not only the development of an AI
tool but also a practical user interface (UI). But it applies equally
2.4.4. Algorithm Domain to the validation side, users can never be forced to use your tool,
they need to want to use it. Getting to this point requires many
Model development in the Build phase is tightly coupled to cycles of user interviews and testing. In Build phase, it is still
issues with data and clinical fit. With access to a larger data possible that the true nature of the clinical need has been
set, priority should be set on maintaining the predictive value, misinterpreted by the development team, or that the proposed
despite potential differences between the sources. It is not technical solution is not capable of attaining the required level
uncommon that the initial technology is unsuited to accommo- of performance. Ideally, however, this phase ends with a product
dating heterogeneous data. Thus, it happens frequently that the which is ready for medical device certification and efficacy
models need to be rebuilt from scratch in this phase, often with studies and all major risks for the Launch phase have been kept
the assistance of more experienced engineers. as low as possible.
Associated with predictive value, the single biggest issue in the
Build phase is the mitigation of bias. The developed models must 2.5. Launch Phase
maintain predictive value on all real-world data (RWD) which the
future product will encounter and not just on a subpopulation. The last phase, Launch, can take up to 3 years depending on the
Patients from different regions can exhibit very different disease type of medical AI product. In this phase, medical device certifi-
patterns, for example. In image generating fields such as cation is achieved and clinical validation is performed.
radiology/pathology, different hardware (scanners/digitalization
software) and different clinical practices (sequence settings/ 2.5.1. Clinical Domain
staining practices) can lead to tremendous differences in image
statistics. Thus, funders and founders should not be blinded by The final clinical validation is the proof that the product fulfills its
high predictive values on single-source data. Models developed promise in the clinical real-world setting. There are a few cases
under such conditions can fail miserably when confronted with where retrospective data (also referred to as real-world evidence
even slightly different data sets. [RWE]) might be used to “prove” clinical products. There are clin-
From an engineering point of view, the finished product will ical, regulatory, and even statistical reasons[39] to argue that these
be deployed somewhere and must fit into the clinical workflow. cases will always prove to be a very small minority (see also
Choosing algorithms which both scale appropriately and are Section 3). In general, then, it will typically be necessary to run
deployable in the context is vital. Algorithm designers should also a clinical trial of any new AI-based medical product. The goal
bear in mind extensibility, e.g., a coherent product roadmap of this trial will be to prove not only safety but also clinical efficacy.
might incorporate plans for future data streams enabling poten- This step is crucial. Without proven clinical efficacy—either for
tial extended applications. clinical benefit or for cost reduction—there is no real incentive
The Build phase is also the phase where the necessity of for customers to buy the product. Thus, the successful completion
explainable AI (xAI) deserves to be considered. Applications, of such a trial is usually the final step before widespread sales of
such as clinical decision support, might require explanations the product in the medical-product marketplace can happen.
as the users (doctors) demand them[36] or are required to use It is crucial to understand that such a trial will almost always
them.[37] A more comprehensive product vision might even have to be a randomized controlled trial (RCT). Only RCTs can
leverage xAI to provide different levels of explanation to different provide the necessary evidence level which is required to change
classes of users (e.g., doctors vs nurses vs patients). For the devel- clinical practice and to find their way into clinical standard-
opers, it might be prudent to utilize xAI to identify potential of-care guidelines. RCTs can also be designed to allow the iden-
hidden prediction biases. For example, it could be that a certain tification of effects on subpopulations, which is often necessary
pathology is coupled to a certain hardware characteristic in the for AI-based products. A major risk for the Launch phase is to
data. In that case, the model might learn to identify that underestimate the time and costs required for the planning and
hardware-tag instead of the pathology.[38] Most importantly, there conduct of an RCT trial, which can amount to several years.
are signs that xAI might be required by medical device regulatory Any evidence established by the trial will be subsequently exam-
authorities in the future.[26] ined under the lens of the trial design. A trial closer to real-world
conditions will bring a higher value to the product (and will be
2.4.5. Outlook necessary if the route to market requires changes to standards
of care); but it runs a higher risk of failure due to uncontrolled
Build phase is the phase which most closely corresponds to the conditions. This risk is mitigated by practices in the previous
public perception of product development. Here, clinical valida- phases working with heterogeneous data sets and minimizing
tion should ideally switch from identifying the clinical need to bias. To ensure an optimal trial outcome, it is highly advisable
fulfilling it. It is also the phase in which most companies stag- to allocate necessary funding and to work as early as possible
nate. Reasons for that can be discoveries which indicate that ini- with a contract research organization (CRO). Particular attention
tial assumptions about data were not accurate. Software solutions should be paid as to whether clinicians, in the trial, actually use the
may be found to contain bugs or are more difficult to build than software in the manner expected by the designers. Once the

Adv. Intell. Syst. 2020, 2, 2000052 2000052 (9 of 14) © 2020 The Authors. Published by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
www.advancedsciencenews.com www.advintellsyst.com

product is approved and it appears on the market, further valida- model failure, design failure can also be the cause. Extensive
tion and changes will be extremely costly. testing in Build phase, both on the model development and the
UI/UX side, should help to keep the risk of such an outcome as
2.5.2. Regulatory Domain low as possible. That said, AI models and software solutions
around them fail at all stages of development. Typically, in
In many senses, the Launch phase is the regulatory phase. In it, Launch, this will occur due to some previously unrecognized reg-
the team will focus on the implementation of a fully compliant ularities in in-house data (“house effect”) which does not transfer
software development process and all steps necessary for medical to RWD. Assuming such a catastrophic failure does not occur,
certification will be taken. The major regulatory risk for this the algorithmic focus should be on how well the model transfers
phase is that key assumptions about the certification process to out-of-house data. Often, despite all previous efforts, RWD are
were wrong and the certification process is massively delayed. still more noisy and less homogenous than previously acquired
Importantly, this phase will also involve considerable efforts to data. The Launch phase is the last phase in which it is possible to
consider the regulatory aspects of a product postlaunch when easily detect lower performance on lower quality data and make
it is already in the market. Processes must be developed to handle final adjustments to the algorithms. To build a product which
surveillance of how the product behaves in real-world usage. If will reach market acceptance, it is important to be prepared
users report a bug, which may cause harm to life (i.e., an adverse for this challenge, to embrace it and to solve it. The team should
event), product managers must already have contingency plans be experienced enough by now.
in place as to how this situation will be handled. Given the cur-
rent state of medical device certification, negotiations must also 2.5.5. Outlook
be entered into with regulatory bodies as to how minor and bug-
fixing software updates will be handled without requiring a rerun For a general nonmedical digital startup, the third phase is the
of clinical trials. A software solution which requires a complete one in which advice turns toward rapidly expanding the user base
clinical trial for every minor bug fix will not be worth much. and optimizing the product–market fit. In a medically regulated
By addressing these issues before they arise, regulators will be product, however, the third phase must precede rapid growth
more willing to put in place framework agreements laying out because the product needs to become certified and clinically vali-
conditions for automatic certification of updates. dated. The processes for certification and clinical validation
remain, to a certain degree, very much the last chance to get
2.5.3. Data Domain product–market fit aligned. Minimizing the risks in the previous
phases of development and working with professional partners
The Launch phase is, in most cases, the final phase in which the both for regulation and clinical validation will keep the risks for
development team can expect to see new forms of medical data. failure low in this final phase.
For certain cases, clinical trials can be constructed such that the
internal team are, subsequently, allowed to train their models on 2.6. Postmarket
the newly acquired data. Retraining the model will likely require
backtesting, demonstrating that, had the new model been used After completing Launch, the product development roadmap
throughout the trial then the algorithm behavior would have been returns to that more typically seen in nonregulated markets.
identical. What happens, however, in the case where backtesting A certified medical AI product has been shown to be safe. If it
indicates minor differences in treatments is still an open debate in has been through an RCT, it may also demonstrate certain stand-
regulatory circles. Notwithstanding, a major risk is that the data ards of clinical efficacy. In most cases, this means that a sales
used in the previous stages were not heterogeneous enough process can proceed and the delivery of the product massively
and RWD within the clinical validation are much noisier or so dif- scaled-up. The addition of new product features will usually
ferent that the predictive value does not hold. It is also paramount follow a slightly condensed version of the Form–Build–Launch
to ensure that the data collected in the clinical validation study cycle. For the product in the market, three topics remain of con-
actually represent the real-world setting. Here, it will be necessary siderable ongoing concern: software updates, data practices, and
to obtain proof for this claim. It should be kept in mind that the surveillance.
acquisition of patient data directly from postlaunch usage of the Software, as it is developed today, is incredibly buggy. Even
product is likely to be rare. It is true that a product which, through when following advanced software development best practices
its usage, accretes additional knowledge about patients is an (e.g., B-Method[40]) software errors persist and must be corrected
incredibly valuable product design. However, patient privacy for. In physical medical devices, the electronic solutions follow
practices throughout the world vary to such an enormous degree decades of engineering-led best practices. In the event of a prob-
that we find it unlikely that many products will successfully follow lem, the scale of the problem (risk) is analyzed and regulators
this path. Thus, data collection in this final premarket phase must usually influence the decision as to whether to recall the product
be planned and performed with the highest scrutiny. or merely fix it in future models. Software combines the ability to
update the working system remotely, at almost zero cost, with a
2.5.4. Algorithm Domain lower quality of initial engineering. Because software for medical
purposes has seen relatively little development to date, no clear
The greatest risk in the Launch phase is that the final software working regulatory practice has yet emerged. What is clear is that
solution will fail. It should be mentioned that at this stage, next to medical products are evaluated in terms of risk: What risk does

Adv. Intell. Syst. 2020, 2, 2000052 2000052 (10 of 14) © 2020 The Authors. Published by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
www.advancedsciencenews.com www.advintellsyst.com

this product pose to the patient? What if it goes wrong? What The presented framework is an original and unique analysis of
must happen if it goes wrong? This evaluation is conducted at the development of AI healthcare products. It divides the devel-
the macrolevel of the whole product, at the microlevel of the indi- opment into three consecutive phases: Form, Build, and Launch.
vidual components, and at the intermediate level of the interface Each phase is set to fulfill important requirements and delever-
which brings the components together. Development teams age the major risks of the subsequent phases. These three phases
need to stay on top of shifting standards of best practice represent distinct stages as they appear in the course of the devel-
and, in all likelihood, should consider medical AI development opment of AI in healthcare products. Researchers, founders, and
to require similar levels of oversight as that historically other stakeholders should be able, with the aid of our framework,
required by aeronautical, spaceflight, and automobile software to easily identify the current phase and relative progress of a
development. project. Within these three phases, the framework covers the
A further, often overlooked process, in developing a medical four most important domains of such a project, namely the clin-
AI solution is the question of shifting clinical standards. AI ical, regulatory, algorithmic, and data domains. Each domain
experts refer to this problem as “stationarity of the data.” For must meet strict requirements to successfully complete the
example, the diagnostic criteria, and how they are encoded in phase. If one of the areas is lagging behind, this will lead to the
medical records, change over time. This is an expert-driven pro- accumulation of—potentially fatal—risks for the project and
cess which trickles down through medical practitioners. The jeopardize the product.
development of a medical AI usually happens in a concentrated We are not aware of other frameworks focusing on AI in
window in time, during which medical practices might be rela- healthcare product development to date. Our approach largely
tively stable, but the deployment of such a solution must then follows the “decision perspective” model introduced by
be used over a much longer time period. Detection, whether Krishnan and Ulrich.[30] Its advantage is that it does not rely
by automated means or by internal human-driven processes, on the perspective of a single domain, but rather divides the
of changes in medical practice must be planned for and pro- development process in a series of decisions. Thus, for each
cesses established as to how to update the software in response. phase, the right decisions can be made to achieve specific goals.
This is a huge challenge, as the whole process of Form, Build, This, of course, paves the way for future research work, which
and Launch might be based on a single snapshot of clinical prac- can focus on more specialized areas of AI in healthcare develop-
tice. A major shift in clinical practice can invalidate most of the ment such as marketing, organizations research, and other
work done during development. topics. Importantly, the advantage of a decision perspective
Surveillance is a concept which healthcare professionals and framework is the focus on what to do and not how to do it.
pharmaceutical manufacturers understand reasonably well. It is Thus, frameworks focusing on how to best perform development
not something which software and AI developers will naturally are complementary to our framework as well as field-specific
consider without prompting. There must be feedback methods roadmaps aimed at the individual challenges to adopt AI.[41]
for reporting misapplications of the software in the wild. The The literature about this topic is too broad to be covered fully
spectrum of events which must be accounted for ranges from here. We will give a short overview of some frameworks in
actual medical adverse events, to users accidentally misunder- use. A key discovery, of the past 20 years, was the coining of
standing certain display elements. Most likely, these methods the concept of an “effectual entrepreneur.”[42] Such an entrepre-
must span the spectrum from paper-based reporting methods neur works at the extreme end of low inherent knowledge—
to in-device automated monitoring. Furthermore, awareness unplannable processes. Over time, they develop an enterprise
must be paid to the fact that streaming of telemetry is usually which increases inherent knowledge and beds down processes,
frowned upon in medical circles; it is too easy to leak patients’ which ultimately lead to a product. In startup culture, such an
private medical data. We are not aware of good examples of med- entrepreneur is characterized using product life-cycle approaches
ical AI device surveillance solutions which do not require ongo- such as lean startup or agile (not to be confused with Toyota’s
ing access to patient medical records and extensive use of clinical lean manufacturing).[32,43] Larger industries, attempting to learn
review panels. No matter what the solution arrived at, it is the from startups have adopted Design Thinking[44]—an essentially
company producing the medical AI device which will be respon- two-step process which begins by focusing on need before moving
sible for product surveillance and must pay for ongoing support to lean startup-style product development. From a funding point of
in this domain or face a loss of medical certification. view, the Business Model Canvas[45] has proved invaluable in
identifying the product–market fit for new companies. Finally,
the Three Lenses of Innovation—a potential product must be
3. Discussion desirable, feasible, and viable—has proven useful to examine
prospective market needs and product–market fit.[46] Our frame-
Healthcare systems all over the world face the same challenges. work is both compatible with and extends these frameworks.
There is an urgent need to deliver better quality healthcare and Importantly, within the framework, we deliberately did not
simultaneously lower costs to keep our healthcare systems sus- focus on funding schemes. The reason here is that current fund-
tainable. AI has the potential to fulfill both of these requirements. ing schemes often lead to a misalignment of interests. The rea-
Although AI often hits the headlines with high performance sons for this status quo are complex. Current funding schemes
in concept studies, certified and clinically validated products for AI in healthcare products—often termed Digital Health—
are still rare. To battle this translational gap, we present a frame- usually follow experiences from products outside of healthcare,
work addressing best practices for the development of AI health- derived in the modern digital era, where a sufficient return of
care products. investment (ROI) is expected within a few years. However,

Adv. Intell. Syst. 2020, 2, 2000052 2000052 (11 of 14) © 2020 The Authors. Published by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
www.advancedsciencenews.com www.advintellsyst.com

despite being digital software products, AI in healthcare products space are typically subject to subsequent regulatory approval.
are more closely related to biotech products sharing the complex In addition, the ethical implications of potentially endangering
development and heavy regulatory requirements. Why then are patients’ health in a trade-off for faster development are tremen-
funding schemes from biotech not adapted to the AI in health- dous. Instead of reducing regulations, we suggest a more mature
care sector? This can be attributed, in part, to the fact that the approach to clinical validation based on realistic approaches such
main exit for biotech companies is a combination of initial public as that contained in our framework. That said, there are alterna-
offering (IPO) and merger and acquisition (M&A),[47] usually in tives. We see huge scope for the certification of AI “platforms” or
partnership with large biotech or pharma companies. In general, manufacturers where only the new components or the manufac-
outside of biotech, IPO activity has significantly decreased;[48] turer need to be certified upon upgrades.[50]
meanwhile, few large acquiring companies have emerged for A related narrative is the increasing discussion of RWD as the
medical AI. Thus, it is very hard for investors to ensure that a real goldmine for AI in medicine. Currently, efforts are being
potentially very large investment sum will yield the required made to allow RWE for the clinical validation of healthcare prod-
ROI. Relatively few investors are even aware of this challenge. ucts.[39,41,51] What is the challenge of RWD for clinical validation
Founders, whether unscrupulous or naive, attract investment with regards to AI in healthcare products? Importantly, the term
by advertising their product niche as being equal to any other “AI” only describes the method used in the product, not the
(AI) software. This strategy, typically involving categorising the model of medicine which lies behind it. Here, a significant obser-
product as a nonmedical device, leads to stark consequences in vation is that many AI solutions developed and marketed today
terms of self-limiting the potential capabilities of the product or, can be considered—often unbeknownst to their creators—
worse, risking subsequent regulatory backlash. Thus, AI in precision medicine approaches.[52] Precision medicine is a type
healthcare development is forced into following development of personalized medicine aiming to integrate all available data for
paths of simpler products due to unreasonable expectations. This each patient into an individual profile and tailoring diagnostics or
inevitably leads to founders cutting corners to pretend fast agile therapy to this profile. AI is a great fit for precision medicine due
development. Evidently, such endeavors are bound to fail, when to its inherent ability to utilize high-dimensional data to identify
the hard requirements of medical device regulation and clinical subpopulations and make individualized predictions.[53] Thus,
validation cannot be met. In the words of venture capitalists, RWD often allows the development of AI products—in the
the main challenge for AI in healthcare startups is the so-called Form and Build phases—as, currently, only RWD has available
“valley of death,” where the startup already exists but does not data for subpopulations. In this sense, RWD can be seen as a
generate revenue. This phase is usually equivalent to the driver of AI development. Data from classic randomized clinical
Build þ Launch phases in our framework. Thus, there is a need trials (RCT), typically lack the granularity to distinguish differing
for an honest appraisal of the development roadmap for AI populations and have—often—less value for AI in healthcare
healthcare products. Here, our framework is an important products. RWD is per definition retrospective and thus has—
contribution. Early-stage funding needs to be made aware of all widely known and accepted—drawbacks, e.g., bias, lack of
the true roadmap and timelines, and expectations need adjust- controls, confounding, and others.[54] Here, the main challenge
ment accordingly. As an example of increased public funding, is that availability is not a valid surrogate for the required level of
a scheme specifically targeting the valley of death is currently evidence for safety and efficacy. One of the greatest hidden risks
set up by the European Union launching in 2021.[49] As the sector in a medical AI product developed on RWD is in algorithmically
matures, we hope that there will be an expansion of the acquisi- enforcing bias. Also, indeed, commercial products in healthcare
tion activities by large pharma- and med-tech companies leading have already been shown to contain considerable racial bias.[55]
to a clearer pathway to product and company development. RCTs, in contrast, have been developed to minimize bias and
The previously identified crisis of the funding models for AI confounding, and provide the highest level of evidence. It is per-
in healthcare products opens up the question as to how AI in fectly possible to perform an RCT to show efficacy and safety also
healthcare products will be funded in the future. The current for subpopulations and hence for any AI in healthcare product.
industry response which we have identified is to lobby for The main difference is that a subpopulation RCT is very costly
reduced regulations. The proposed rationale for this is that faster and time-consuming, as a sufficient number of patients per sub-
and less-regulated technological innovation will ultimately group need to be enrolled. A push to use RWE instead of RCTs
deliver on the healthcare goals of broadening access while reduc- for the validation of AI in healthcare products is essentially a
ing costs. This cost argument has led to a willingness, on the part push to trade the level of evidence of safety and efficacy for lower
of regulators, to open up avenues for clearance of medical soft- costs, increased speed of development and higher risk for
ware solutions without the full burden associated with traditional patients due to bias and confounding. It is fair to say that such
pharmaceutical drug and medical device regulations. This open- a push would ethically and publicly never be acceptable if it was
ing of alternative regulatory routes is tempered by an awareness suggested for pharmaceutical products. Thus, it is hard to justify
that medical AI can, and will, directly impact on patient health. such a move for AI in healthcare products which often might
In practice, however, the reduction of regulatory barriers appears have the same potential for patient risk. Consequently, our
to lead to a potential diversion of spending, from clinical valida- framework stresses the importance of an RCT as the required
tion and software engineering, to paying experts on the naviga- level of clinical validation for the overwhelming majority of AI
tion of such shortcuts—and subsequent enormous payment for in healthcare products in the Launch phase.
product-risk insurance when the product reaches market. We do In developing our framework, we have largely envisaged
not see this as a sustainable path, indeed much of the current medium- to high-risk medical AI products. This has streamlined
excitement around health-tech fails to appreciate that sales in this our thought processes and allowed us to develop rules

Adv. Intell. Syst. 2020, 2, 2000052 2000052 (12 of 14) © 2020 The Authors. Published by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
www.advancedsciencenews.com www.advintellsyst.com

particularly of use to a medical audience. However, we do not see [6] Z. Zhang, Y. Zhao, A. Canes, D. Steinberg, O. Lyashevska, Ann. Transl.
any real differences in the process for developing low-risk prod- Med. 2019, 7, 152.
ucts. The only place where there is a question mark is in products [7] L. Breiman, J. Friedman, C. J. Stone, R. A. Olshen, Classification and
which attempt to fall below all risk thresholds for medical regu- Regression Trees, Chapman and Hall/CRC, Boca Raton, FL 1984.
lation. At this point, we would question the wisdom of any strat- [8] V. Podgorelec, P. Kokol, B. Stiglic, I. Rozman, J. Med. Syst. 2002,
egy which looks at providing clinically oriented benefits while 26, 445.
[9] P. McCullagh, J. A. Nelder, Generalized Linear Models, Chapman and
attempting to market itself as a nonmedical product. This con-
Hall/CRC, Boca Raton, FL 1989.
tinues to run the risk of falling foul of regulatory attention while
[10] C. Cortes, V. Vapnik, Mach. Learn. 1995, 20, 273.
also failing to attain the enhanced sales values associated with [11] A. Mammone, M. Turchi, N. Cristianini, WIREs Comput. Stat. 2009,
medical products. In any case, a familiarity with our framework 1, 283.
will allow key decision-makers to evaluate the potential costs and [12] C. E. Rasmussen, C. K. I. Williams, Gaussian Processes for Machine
benefits of following a regulatory path. Learning, MIT Press, Cambridge, MA 2006.
[13] M. Seeger, Int. J. Neural Syst. 2004, 14, 69.
[14] P. Patel, A. Thakkar, Int. J. Electr. Comput. Eng., 2020, 10, 538.
4. Conclusion [15] M. E. Peters, M. Neumann, M. Iyyer, M. Gardner, C. Clark, K. Lee,
L. Zettlemoyer, Proc. of the 2018 Conf. of the North American
We have developed a framework which outlines, what we con- Chapter of the Association for Computational Linguistics: Human
sider to be, best practices in AI in healthcare development. Language Technologies, Vol. 1 (Long Papers), Association for
Our goal is to pave the way for products which are both effica- Computational Linguistics, New Orleans, LA 2018, pp. 2227–2237.
cious and safe for patients. Through direct participation in the [16] J. Saha, C. Chowdhury, S. Biswas, Deep Learning Techniques for
development of a spectrum of medical AI products, our experi- Biomedical and Health Informatics (Eds: S. Dash, B. R. Acharya,
ence, and that of our peers, is of enormous waste in the process. M. Mittal, A. Abraham, A. Kelemen), Springer International
Current initiatives to improve the flow of AI in healthcare prod- Publishing, Cham 2020, pp. 101–126.
ucts aim at reducing costs and speeding up development through [17] S. Rendle, F. Machines, in Proc. IEEE Int. Conf. on Data Minning, IEEE,
means which potentially increase risks for patients. Our frame- Sydney 2010.
work outlines an alternative to this approach. We aim to improve [18] R. Ying, R. He, K. Chen, P. Eksombatchai, W. L. Hamilton, J. Leskovec,
in Proc. 24th ACM SIGKDD Int. Conf. on Knowledge Discovery and Data
the development situation of AI-driven medical products by
Mining, Association for Computing Machinery, New York, NY 2018.
enhancing accountability, transparency, and planning, hopefully
[19] A. Panesar, Machine Learning and AI for Healthcare: Big Data for
increasing success rates for such products, without facing some
Improved Health Outcomes (Eds: A. Panesar), Apress, Berkeley, CA
of the safety and ethical trade-offs mentioned previously, to fulfill 2019, pp. 255–304.
the promise of AI in healthcare: better care at lower costs. [20] J. A. Savino, R. Latifi, in The Modern Hospital: Patients Centered,
Disease Based, Research Oriented, Technology Driven (Eds: R. Latifi),
Springer, Cham 2019, pp. 375–387.
Acknowledgements [21] M. Livne, XPOMET©: 360 Next Generation Healthcare, MWV
We would like to acknowledge Michelle Livne, Nicholas Borsotto, Medizinisch Wissenschaftliche Verlagsgesellschaft, Berlin, Germany
Tim Higgins, and Dorothée Doepfer for valuable ideas and proofreading. 2019.
[22] E. Basch, A. M. Deal, A. C. Dueck, H. I. Scher, M. G. Kris, C. Hudis,
D. Schrag, JAMA 2017, 318, 197.
[23] A. Hosny, C. Parmar, J. Quackenbush, L. H. Schwartz,
Conflict of Interest
H. J. W. L. Aerts, Nat. Rev. Cancer 2018, 18, 500.
The authors declare no conflict of interest. [24] N. Larios Delgado, N. Usuyama, A. K. Hall, R. J. Hazen, M. Ma,
S. Sahu, J. Lundin, Npj Digital Med. 2019, 2, 1.
[25] T. Davenport, R. Kalakota, Future Heal. J. 2019, 6, 94.
[26] S. Reddy, J. Fox, M. P. Purohit, J. R. Soc. Med. 2019, 112, 22.
Keywords [27] E. Racine, W. Boehlen, M. Sample, Healthc. Manage. Forum 2019,
artificial intelligence, digital health, healthcare, innovation frameworks, 32, 272.
medicine [28] D. Ben-Israel, W. B. Jacobs, S. Casha, S. Lang, W. H. A. Ryu,
M. de Lotbiniere-Bassett, D. W. Cadotte, Artif. Intell. Med. 2020,
Received: March 24, 2020 103, 101785.
Published online: July 2, 2020 [29] D. W. Kim, H. Y. Jang, K. W. Kim, Y. Shin, S. H. Park, Korean J. Radiol.
2019, 20, 405.
[30] V. Krishnan, K. T. Ulrich, Manage. Sci. 2001, 47, 1.
[1] Lifetime Risk of Developing or Dying from Cancer, https://www. [31] B. Masters, P. Thiel, Zero to One: Notes on Start Ups, or How to Build
cancer.org/cancer/cancer-basics/lifetime-probability-of-developing- the Future, Virgin Books, London, UK 2015.
or-dying-from-cancer.html (accessed: March 2020). [32] E. Ries, The Lean Startup: How Today’s Entrepreneurs Use Continuous
[2] The GBD 2016 Lifetime Risk of Stroke Collaborators, New Engl. J. Innovation to Create Radically Successful Businesses, Currency,
Med. 2018, 379, 2429. New York, NY 2011.
[3] F. Griffiths, M. Ooi, IEEE Instrum. Meas. Mag. 2018, 21, 29. [33] W. J. Gordon, A. Landman, H. Zhang, D. W. Bates, Npj Digital Med.
[4] A. Esteva, A. Robicquet, B. Ramsundar, V. Kuleshov, M. DePristo, 2020, 3, 1.
K. Chou, C. Cui, G. Corrado, S. Thrun, J. Dean, Nat. Med. 2019, 25, 24. [34] B. Englich, T. Mussweiler, F. Strack, Pers. Soc. Psychol. Bull. 2016,
[5] M. Leshno, V. Lin, A. Pinkus, S. Schocken, Neural Netw. 1993, 6, 861. 2, 188.

Adv. Intell. Syst. 2020, 2, 2000052 2000052 (13 of 14) © 2020 The Authors. Published by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
www.advancedsciencenews.com www.advintellsyst.com

[35] O. Bergman, T. Ellingsen, M. Johannesson, C. Svensson, Econ. Lett. [45] M. C. León, J. I. Nieto-Hipólito, J. Garibaldi-Beltrán, G. Amaya-Parra,
2010, 107, 66. P. Luque-Morales, P. P. Magaña-Espinoza, J. Aguilar-Velazco, J. Med.
[36] E. H. Shortliffe, M. J. Sepúlveda, JAMA 2018, 320, 2199. Syst. 2016, 40, 144.
[37] P. Hacker, R. Krestel, S. Grundmann, F. Naumann, Artif. Intell. Law [46] How to prototype a new business, https://www.ideou.com/blogs/
2020, https://doi.org/10.1007/s10506-020-09260-6. inspiration/how-to-prototype-a-new-business (accessed: March 2020).
[38] J. R. Zech, M. A. Badgeley, M. Liu, A. B. Costa, J. J. Titano, [47] Silicon Valley Bank, 2019 annual healthcare report, https://www.svb.
E. K. Oermann, PLoS Med. 2018, 15, e1002683. com/globalassets/library/managedassets/pdfs/healthcare-report-
[39] J. Wilkinson, D. R. Brison, J. M. N. Duffy, C. M. Farquhar, 2019-annual.pdf (accessed: March 2020).
S. Lensen, S. Mastenbroek, M. Wely, A. Vail, Hum. Reprod. 2019, [48] X. Gao, J. R. Ritter, Z. Zhu, Financ. Quant. Anal. 2013, 48, 1663.
34, 2093. [49] E. U. Plans, $3.9 billion fund for startups in ‘valley of death’, https://
[40] S. Schneider, The B-Method, Red Globe Press, Basingstoke 2001. www.bloomberg.com/news/articles/2019-11-23/eu-plans-3-9-billion-
[41] B. Allen, S. E. Seltzer, C. P. Langlotz, K. P. Dreyer, R. M. Summers, fund-for-startups-in-valley-of-death (accessed: March 2020).
N. Petrick, D. Marinac-Dabic, M. Cruz, T. K. Alkasab, R. J. Hanisch, [50] F. D. A. Pre-Cert, http://www.fda.gov/medical-devices/digital-health/
W. J. Nilsen, J. Burleson, K. Lyman, K. Kandarpa, J. Am. Coll. Radiol. digital-health-software-precertification-pre-cert-program (accessed:
2019, 16, 1179. March 2020).
[42] S. D. Sarasvathy, Acad. Manage. Rev. 2001, 26, 243. [51] N. A. Dreyer, Ther. Innov. Regul. Sci. 2018, 52, 362.
[43] K. N. Manjunath, J. Jagadeesh, M. Yogeesh, Int. Mutli-Conf. on [52] M. R. Kosorok, E. B. Laber, Annu. Rev. Stat. Appl. 2019, 6, 263.
Automation, Computing, Communication, Control and Compressed [53] D. D. Miller, E. W. Brown, Am. J. Med. 2018, 131, 129.
Sensing, IEEE, New York, USA 2013, pp. 26–34. [54] J. M. Franklin, S. Schneeweiss, Clin. Pharmacol. Ther. 2017, 102, 924.
[44] J. P. Roberts, T. R. Fisher, M. J. Trowbridge, C. Bent, Healthcare 2016, [55] Z. Obermeyer, B. Powers, C. Vogeli, S. Mullainathan, Science 2019,
4, 11. 366, 447.

Adv. Intell. Syst. 2020, 2, 2000052 2000052 (14 of 14) © 2020 The Authors. Published by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim

You might also like